disaggregated inference thoughts: what are you seeing?

[localpatch78]

Posting this in case it saves someone else a little time.

I'm helping a small team set up a local workflow. We don't need a giant stack, but we do care about tokens/sec.

Would rather learn from your scars than collect my own all at once.

For context, I'm not chasing leaderboard numbers. I mostly want the system to feel consistent day to day.

The part that surprised me was how much the boring details mattered more than the model choice.

[bytestack]

Not sure I agree there. I get why that seems right, but on Threadripper box the bottleneck felt elsewhere.

[papermixer4]

Ran a quick check on my box. My result was mixed: first run looked bad, second run looked normal. At this point the real bottleneck is my patience.

[vramport]

I tested this tonight. I couldn't reproduce the slowdown until I pushed context much higher. The main thing that moved was latency, not quality. Turns out 'just one more test' was a lie.

[vrambandit]

I tested this tonight. I couldn't reproduce the slowdown until I pushed context much higher. The main thing that moved was latency, not quality. Turns

I'd approach it in a boring order. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at HBM constraints. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[papermixer4]

That matches what I was seeing. The pattern around prefix caching has been pretty similar for me.

[edgeriver]

Thanks, that helps. That matches what I saw on 7900 XTX. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable. At this point the real bottleneck is my patience.

[localpatch]

Interesting, I hadn't tried that. I don't think your conclusion is far off. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

[vramport]

I'd approach it in a boring order. Before rebuilding anything, I'd test one clean baseline with Llama and compare from there. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[kernelrunner]

Thanks, that helps. I ended up in about the same place using Ollama. The biggest change for me was simplifying the stack instead of adding one more optimization layer.